DeepSeek Drops V4 Preview – and Turns Up the Heat in the AI Race

Bloomberg, CNBC, Reuters, AlJazeera, the New York Times contributed to this report.

China’s AI upstart DeepSeek is back with a new model – and a familiar strategy: fast, cheap, and open to everyone.

On Friday, the Hangzhou-based company rolled out a preview of its long-awaited V4 large language model, giving developers early access to test what it can do. Like its earlier releases, V4 is open-source, meaning users can download it, tweak it, and run it on their own systems.

Two versions are on offer. A heavier “pro” model aimed at deeper performance, and a lighter “flash” version built for speed and efficiency. DeepSeek says both are tuned for agent-style tasks – the kind of AI that can plan, reason, and execute multi-step actions – while keeping inference costs down.

That last part matters. Running AI models isn’t cheap, and DeepSeek is making a habit of undercutting rivals.

The company already rattled the industry with its R1 model, which landed in early 2025 and punched well above its weight. It matched – and in some cases beat – leading systems from OpenAI and Google, despite being built in just a couple of months and reportedly costing under $6 million. That claim raised eyebrows, but the impact was real. Investors suddenly had to consider a world where cutting-edge AI didn’t require Silicon Valley-sized budgets.

V4 is unlikely to cause the same shockwave. The market has adjusted. Chinese AI is no longer dismissed as second-tier – it’s now expected to be competitive, and often cheaper.

Still, this release lands in a very different environment. Domestic rivals are circling fast. Alibaba and ByteDance have both pushed out new models this year, turning China’s AI sector into a crowded, high-speed contest. That wasn’t the case when R1 first dropped.

The bigger shift might be under the hood. Questions are swirling about what chips power V4. With U.S. export controls limiting access to top-tier hardware from Nvidia, Chinese firms have been forced to adapt. Huawei says its Ascend AI systems can support V4, hinting at a future where Chinese models run increasingly on homegrown tech.

If that holds up, it’s a big deal. Less reliance on foreign chips, more control over the full AI stack – and potentially faster development cycles inside China.

For now, V4 looks like a solid step forward rather than a dramatic leap. Strong performance, lower costs, open access. Enough to keep pressure on global players who are spending billions to stay ahead.

The gap between US and Chinese AI hasn’t vanished, but it’s clearly shrinking. And releases like this make one thing obvious: the race isn’t slowing down.